7. N. Read, S. Sachdev, Phys. Rev. B 42, 4568–4589 (1990).8. G. Murthy, S. Sachdev, Nucl. Phys. B 344, 557–595 (1990).9. N. Read, S. Sachdev, Phys. Rev. Lett. 66, 1773–1776

(1991).10. O. I. Motrunich, A. Vishwanath, Phys. Rev. B 70, 075104

(2004).11. M. A. Metlitski, M. Hermele, T. Senthil, M. P. A. Fisher, Phys.

Rev. B 78, 214418 (2008).12. E. Dyer, M. Mezei, S. S. Pufu, S. Sachdev, J. High Energy Phys.

2015, 37 (2015).13. T. Li, F. Becca, W. Hu, S. Sorella, Phys. Rev. B 86, 075111

(2012).14. S.-S. Gong, W. Zhu, D. N. Sheng, O. I. Motrunich,

M. P. A. Fisher, Phys. Rev. Lett. 113, 027201 (2014).15. A. W. Sandvik, Phys. Rev. Lett. 98, 227202 (2007).16. R. G. Melko, R. K. Kaul, Phys. Rev. Lett. 100, 017203

(2008).17. F. J. Jiang, M. Nyfeler, S. Chandrasekharan, U.-J. Wiese, J. Stat.

Mech. 2008, P02009 (2008).18. A. W. Sandvik, Phys. Rev. Lett. 104, 177201 (2010).19. R. K. Kaul, Phys. Rev. B 84, 054407 (2011).20. K. Harada et al., Phys. Rev. B 88, 220408 (2013).21. K. Chen et al., Phys. Rev. Lett. 110, 185701 (2013).

22. M. S. Block, R. G. Melko, R. K. Kaul, Phys. Rev. Lett. 111, 137202(2013).

23. S. Pujari, F. Alet, K. Damle, Phys. Rev. B 91, 104411(2015).

24. A. Nahum, J. T. Chalker, P. Serna, M. Ortunõ, A. M. Somoza,Phys. Rev. X 5, 041048 (2015).

25. G. J. Sreejith, S. Powell, Phys. Rev. B 89, 014404 (2014).26. A. B. Kuklov, M. Matsumoto, N. V. Prokof’ev, B. V. Svistunov,

M. Troyer, Phys. Rev. Lett. 101, 050405 (2008).27. O. I. Motrunich, A. Vishwanath, http://arxiv.org/abs/0805.

1494 (2008).28. L. Bartosch, Phys. Rev. B 88, 195140 (2013).29. J. M. Luck, Phys. Rev. B 31, 3069–3083 (1985).30. A. W. Sandvik, V. N. Kotov, O. P. Sushkov, Phys. Rev. Lett. 106,

207203 (2011).31. F. Léonard, B. Delamotte, Phys. Rev. Lett. 115, 200601

(2015).32. Supplementary materials are available on Science Online.33. Y. Tang, A. W. Sandvik, Phys. Rev. Lett. 107, 157201

(2011).34. A. Banerjee, K. Damle, J. Stat. Mech. 2010, P08017

(2010).35. Y. Tang, A. W. Sandvik, Phys. Rev. Lett. 110, 217213 (2013).

ACKNOWLEDGMENTS

The research was supported by the National Natural ScienceFoundation of China, grant no. 11175018 (W.G.); the FundamentalResearch Funds for the Central Universities (W.G.); U.S. NSFgrant no. DMR-1410126 (A.W.S.); and the Simons Foundation(A.W.S.). H.S. and W.G. gratefully acknowledge support from theCondensed Matter Theory Visitors Program at Boston University,and A.W.S. is grateful to the Institute of Physics of the ChineseAcademy of Sciences and to Beijing Normal University for theirsupport. Some computations were carried out using BostonUniversity’s Shared Computing Cluster. We thank K. Damle forsuggesting a definition of spinons that uses valence-bond strings.Computer code is available from the authors upon request.

SUPPLEMENTARY MATERIALS

www.sciencemag.org/content/352/6282/213/suppl/DC1Supplementary TextFigs. S1 to S10References (36–41)

22 September 2015; accepted 1 March 2016Published online 17 March 201610.1126/science.aad5007

BRAIN CONNECTIVITY

Task-free MRI predicts individualdifferences in brain activity duringtask performanceI. Tavor,1,2 O. Parker Jones,1 R. B. Mars,1,3 S. M. Smith,1 T. E. Behrens,1,4 S. Jbabdi1*

When asked to perform the same task, different individuals exhibit markedly differentpatterns of brain activity. This variability is often attributed to volatile factors, such as taskstrategy or compliance.We propose that individual differences in brain responses are, to alarge degree, inherent to the brain and can be predicted from task-independentmeasurements collected at rest. Using a large set of task conditions, spanning severalbehavioral domains, we train a simple model that relates task-independent measurementsto task activity and evaluate the model by predicting task activation maps for unseensubjects using magnetic resonance imaging. Our model can accurately predict individualdifferences in brain activity and highlights a coupling between brain connectivity andfunction that can be captured at the level of individual subjects.

We all differ in howwe perceive, think, andact. Our brains also differ in how theysolve tasks. Understanding these indi-vidual differences in brain activity is animportant goal in neuroscience, as it pro-

vides a route for linking brain and behavior.Often, individual differences in brain activity

induced by an experimental task are attributed totwo possible factors. First, they may be accountedfor by differences in gross brainmorphology. Thevastmajority of brain-imaging studies rely on thespatial alignment of different brains (registration)

to account for between-subject discrepancies ingross anatomy (1). Second, subjects may use dif-ferent strategies or cognitive processes that in-volve different brain circuits. Psychologists designtheir tasks with great care to limit this source ofvariability. Nevertheless, individual variationsare seen in all behavioral domains (2–6). Thesebetween-subject differences are often treated as“noise” in imaging studies and discarded throughthe process of averaging individual responses (7).Indeed, it is thought that such individual differ-ences are mainly explained by volatile factors re-lated to the behavior.We investigated the possibility that individual

differences in brain activation are inherent fea-tures of individuals and, to a large degree, inde-pendent of volatile factors.We explored the extentto which individual differences in task-evokedbrain activity can be predicted by differences inthe functional connectivity of the brain, acquiredin a magnetic resonance imaging (MRI) scannerwhile the subjects are at rest and not performing

any explicit task.We aimed to predict several task-evokedactivitymapsmatching individual subject’smaps in multiple behavioral domains based on asingle task-free scan (i.e., unconstrained cognitionduring rest, with no explicit experimental task)of any given subject.We designed a set of regression-based models

that use task-independent features to predict in-dividual task-evokedresponses. Following thehypo-thesis that functional differentiation in the braincan be understood in terms of the underlyinglong-range brain connections and interactions(8, 9), we used predictors based on functional con-nectivity at rest (10). Brain networks, extractedfrom resting-state data sets, qualitatively resem-ble task-evoked networks at the group level (11).We therefore hypothesized that, using functionalconnectivity at rest, we could predict individualvariations in task responses.We also used predic-tors encoding individual brain morphology (grossstructure) andmicrostructure to yield a total of107 predictors [seeMaterials andMethods in sup-plementarymaterials (SM)]. All of our predictorswere based on imaging subjects at rest and wereindependent of any given task. The model wastrained to map between the predictors and taskactivations in a cohort of subjects for each task,and subsequently, the trainedmodel was appliedto out-of-sample (unseen) subjects to predict theirtask activations (a leave-one-out approach, seeMaterials andMethods in SM for details). Through-out, we focused on predicting task activations onthe cortex, although the approach can easily beextended to incorporate subcortical gray matter.Our data were a subset (98 subjects) of the

Human Connectome Project (HCP) database(12). HCP data were chosen for their inclusion ofresting-state measurements, diffusion-weightedMRI, and structural MRI, as well as task-evokeddata spanning several behavioral domains. Wecould therefore use the same set of task-free datato test predictions of several different tasks. TheHCP task data include seven behavioral domains,and each task set comprises several statisticalmaps pertaining to different aspects of each task

216 8 APRIL 2016 • VOL 352 ISSUE 6282 sciencemag.org SCIENCE

1Oxford Centre for Functional Magnetic Resonance Imagingof the Brain (FMRIB), Nuffield Department of ClinicalNeurosciences, John Radcliffe Hospital, Oxford OX3 9DU, UK.2Department of Diagnostic Imaging, Sheba Medical Center,Tel Hashomer, 52621, Israel. 3Donders Institute for Brain,Cognition and Behaviour, Radboud University Nijmegen, 6525EZ Nijmegen, Netherlands. 4Wellcome Trust Centre forNeuroimaging, University College London, London, WC1N3BG, UK.*Corresponding author. E-mail: saad@fmrib.ox.ac.uk

RESEARCH | REPORTSCorrected 11 April 2016; see full text.

on

Janu

ary

23, 2

017

http

://sc

ienc

e.sc

ienc

emag

.org

/D

ownl

oade

d fr

om

SCIENCE sciencemag.org 8 APRIL 2016 • VOL 352 ISSUE 6282 217

Fig. 1. Predicting individual variations in task maps. Figure shows actual and predicted thresholded task maps in three subjects and four different taskcontrasts. The model is able to capture striking variations between subjects in the shape, topology, and extent of their activation maps. Predictions andcomparisons with actual maps for all behavioral domains and more subjects are shown in figs. S1 to S3 and movies S1 to S7.

Fig. 2. Specificity of the individual predictions. A subject’s predicted mapis more similar to the subject’s actual map than to the rest of the subjects.(A) Predicted maps (zoomed on the right frontal lobe) of the MATH-STORYcontrast from the LANGUAGE task of three subjects are overlapped with theiractual activation maps (top row). We also overlap the subjects’ predictionswith the activation map of the median subject (bottom row). Blue representsactual activation, red is the predicted activation, and yellow is the overlap.Themaximum overlap is obtained when comparing a given subject’s predictionwith their own actual map. (B) Pearson correlation matrix between actual (col-umns) and predicted activations (rows). The correlation matrix is noticeablydiagonal-dominant, which indicates that, on average, the model prediction for

any given subject is more similar to the subject’s own map than to other sub-jects’ maps.This is also shown as a histogram plot, where the extradiagonalelements of the correlation matrix (subject X versus subject Y) are comparedwith the diagonal elements (subject X versus subject X).The vertical dashedline corresponds to the median of the correlation coefficients along thediagonal. (C) Correlation matrices and histograms for six additional behavioraldomains.When we normalized the rows and columns of the correlation matrices(which removes the mean and accounts for higher variability in actual thanin predicted maps), the diagonal-dominance is even more prominent. In allcases, a Kolmogorov-Smirnov test between the two distributions (self versusother) gives a highly significant difference (P < 10−10).

RESEARCH | REPORTSCorrected 11 April 2016; see full text.

on

Janu

ary

23, 2

017

http

://sc

ienc

e.sc

ienc

emag

.org

/D

ownl

oade

d fr

om

[all tasks and derived parametric maps are de-scribed in (13)]. A total of 47 independent taskmaps (z scores) were available for each subject(see Materials and Methods in SM).We aimed to predict not how subjects activate

on average in a given task but how they differ(from each other) in their activation patterns. Thisaspect of the model is illustrated (Fig. 1) in fourdifferent task conditions (maps thresholded asdescribed in Materials andMethods). The modelcould predict qualitative differences betweensubjects, in terms of shape, position, size, and to-pography of activations. Similar individual vari-ations predicted across more subjects and allbehavioral domains used by the HCP are shownin figs. S1 to S3 and movies S1 to S7. The modelcould precisely predict individual task variationsacross all behavioral domains, on the basis of asingle task-free data set. To get an impressionof how similar our predictions are to the actualtask maps, we overlapped the predictions forthree subjects with their actual activation maps(Fig. 2A) for the contrast MATH-STORY of theLANGUAGE task. For comparison,we also overlap-ped the same subjects with the actual activationmapof a canonical (median) subject. Themaximum

overlap was obtained when comparing a givensubject’s prediction with their actual map. Thiswas despite our use of a leave-one-out approach:The predictive model for subject X had not seenthe activation map of subject X during train-ing, but it had seen themaps of all other subjects.Nevertheless, the prediction was more similar tothe map it had not seen (subject X) than to themaps it had seen (all but subject X). The modeldid not learn what the activations for a given tasklooked like, but rather it learned how tomap fromthe features (resting-state connectivity and mor-phology) to the task maps in individual subjects.To quantitatively assess the performance of

the model, we estimated the spatial correlationbetween the (unthresholded) predictions and ac-tual maps for all pairs of subjects (Fig. 2, B and C).Each entry in the matrix is the Pearson correla-tion between the taskmap of one subject and thepredicted map of another (off-diagonal) or thesame (on-diagonal) subject. The correlationmatrixis noticeably diagonal-dominant, which indicatesthat, on average, the model prediction for anygiven subject is more similar to the subject’s ownmap than to other subjects’ maps (see also thehistograms shown in Fig. 2, B and C, comparing

on-diagonal with off-diagonal entries). Normal-izing rows and columns of the correlationmatrix(which removes the overallmean correlation andaccounts for the fact that the actual maps aremore variable than the predicted maps) showsthe diagonal-dominance evenmore clearly (see alsofigs. S4 to S6, in which we show the same resultsfor all contrast maps). The diagonal-dominanceis apparent for all tasks and contrast maps, withthe exception of one contrast map (GAMBLINGREWARD). The reason is that activations for thiscontrast are restricted to subcortical gray matter,whereas our model only makes predictions forthe cortex. This result, i.e., the diagonal dominanceof the spatial correlation matrix, is nontrivialconsidering that our leave-one-out procedureensures that, whenever a model is applied topredict a given subject, it has never seen thatsubject’s task map during training. Yet, the pre-diction matches that subject’s task data betterthan the subjects that it has seen during training.Themodel’s ability to generalize beyond the train-ing subjects has important implications. It canpredict activations in individuals for which thereare no available task data (e.g., patients who can-not perform the task).

218 8 APRIL 2016 • VOL 352 ISSUE 6282 sciencemag.org SCIENCE

Fig. 3. Capturing qualitative and quantitative interindividual differences. (A) Variations in location, shape, and topology are predicted by themodel (contrast:LANGUAGEMATH-STORY). (B) Peak Z scores were calculated for each hemisphere to examine how well themodel can predict the amount of activation for eachsubject. A lateralization index (difference between right and left peak activation levels) is then calculated for each subject for both predicted and actual data and isshown as red and blue bars, respectively (LANGUAGE task).Themodel is able to predict individual subject’s lateralization index for both contrasts, including the casewhere the majority of the subjects are left-lateralized. Statistical tests: MATH-STORY [correlation coefficient (r) = 0.47, P < 10−5], STORY-MATH (r = 0.48, P < 10−6).

RESEARCH | REPORTSCorrected 11 April 2016; see full text.

on

Janu

ary

23, 2

017

http

://sc

ienc

e.sc

ienc

emag

.org

/D

ownl

oade

d fr

om

Themodel can precisely capture interindividualvariability (table S3). There are different types ofsuch variations. Functional brain areas in individ-ual subjects can greatly vary in size, location, andshape (14). Primary visual cortex, for instance, canvary (across different subjects) by a factor of twoor more in surface area (15). Such variability stillallows one-to-one mapping of brain areas acrosssubjects, provided that individual areas can bemapped in individual subjects. But another typeof variation is topological variability, which forexample means that different subjects may havedifferent patterns of activity with no obvious one-to-onecorrespondencebetweensubjects. Both typesof variability (shape or size versus topological) canbe captured by this model (Fig. 3A). Note thattopological variability cannot be accounted forusing spatial alignment between subjects, andthus, comparing or averaging subjects’ activationswhen they have different topologies is an openproblem. The fact that we can predict such vari-ability from task-free data may enable new strat-egies formatching brain networks, as opposed tobrain areas, across subjects.Our model can also make predictions on the

distribution of activity across different systems,such as in the case of hemispheric lateralization.Lateralization of function is seen commonly in thedomain of language (16) but also in attentionalprocessing (17). Our model can predict such dif-ferences across individuals in a language task, as

shown in Fig. 3B. We estimated a lateralizationindex defined as the difference between righthemisphere and left hemisphere peak Z (aver-aged over a 10-mm radius sphere around thepredicted peak) from the predictions and the ac-tual data. The model is able to precisely predictan individual subject’s lateralization index. Wealso predicted two different contrasts from thesame task (MATH-STORY and STORY-MATH),where the former tends to show an approxi-mately equal distribution among left- and right-lateralized individuals, and the latter shows moreleft-lateralized subjects. The model can capturethe lateralized language trend, even for the mi-nority of right-lateralized subjects, despite the factthat, for the STORY-MATH contrast, it has beentrained on subjects the majority of whom areleft-lateralized.The model could also predict atypical activa-

tion patterns. In Fig. 4, we show in five differentbehavioral domains examples of subjects that ei-ther do or do not conform to the pattern seen inthe average subject. The model predicted activa-tions in regions that were not activated on average.Conversely, themodel correctly predicted the lackof activations in regions that were activated onaverage.Overall, a simple set of models trained to learn

a mapping between task-free and task-evokedmaps could be used to predict individual differ-ences in activation maps accurately across a

wide range of behavioral domains. The predictedmaps matched even large individual variationsin the spatial layout of brain activity, includingposition, shape, size, and topology of functionalactivations (e.g., whether the activity is localizedor spread out, or whether it consists of a singleregion or multiple subregions). The model couldalso predict individual differences in the amountof activation in a given task from a task-freedata set.The model mainly used resting-state connec-

tivity in addition to a few structural features cap-turing local morphology and microstructure(seeMaterials andMethods in SM). On averageacross all tasks, all features participated in thepredictions except for subcortical connectivityfeatures (excluding the cerebellum—see figs. S7and S8). Although the structural features wereexploited by the model (fig. S7), removing themdid not affect the performance of the model (fig.S9), and using a model purely based on structuralfeatures abolishes interindividual variability in thepredictions (fig. S10). This suggests that resting-state connectivity alone is sufficient to predictindividual variability in task maps, independent-ly from variations of morphology as captured byour handful of structural features. It is, however,possible that additional anatomical variabilitymay also be indirectly captured by the resting-state data.Why can we predict task-evoked activity from

resting-state data? A goodmatch between resting-state networks and task networks at the grouplevel has been shown before (11, 18), where it hasbeen argued that functional networks are conti-nuously interacting with each other at rest, withthe same functional hierarchy that is seen duringaction and cognition. Our model provides an ex-plicit mapping from resting-state data to taskmaps that corroborates this assumption, but it isimportant to consider alternative explanations ofour result. Intersubject alignment may be sub-optimal when using anatomy alone. Our model istherefore possibly accounting for functional mis-alignment between subjects (19). However, themodel not only tracks variation in the positionand shape of functional areas but also capturesvariations in the topology of the brain networksactivated in individual subjects and can predictactivations of atypical subjects. Therefore, it isunlikely that the standard approach of aligningsubjects while preserving the brain topologywill be able to fix these types of misalignments.Our model can make quantitative predictions, interms of the amount of activity that individual sub-jects display during a task. Such predictions areunlikely to arise frommisalignment considerations.Resting-state datamay provide ameans for “calibra-tion” of the blood oxygen–level–dependent (BOLD)signal (20). However, our predictors are based onmeasures of connectivity rather than raw BOLDsignal strength per se [although signal-to-noisemay still affect functional connectivity at rest (21)].Determiningwhat information in the resting-statesignal is driving our predictions will be importantfor understanding its nature and potential. Becauseall the predictors that were used in this study are

SCIENCE sciencemag.org 8 APRIL 2016 • VOL 352 ISSUE 6282 219

Fig. 4. Predictions in atypical subjects.The figure demonstrates the ability of our prediction model todetect intersubject variability when subjects differ from the group-averaged activation. In each row, thegroup activation for each behavioral domain is shown on the left, and the actual and predicted activationsfor two subjects are shown on the right. In each row, we show activations and predictions for one subject(A),whose activation is similar to the group activation, and another (B)who differs from it, as shown by theblack circles.Themodel can capture the presence of clusters that are not active in the group (rows 1,4, and5), as well as the absence of clusters that are active in the group (rows 2 and 3). Note that subjects A andBare not the same pair of subjects across behavioral domains.

RESEARCH | REPORTSCorrected 11 April 2016; see full text.

on

Janu

ary

23, 2

017

http

://sc

ienc

e.sc

ienc

emag

.org

/D

ownl

oade

d fr

om

based on scans acquired at rest, our model isblind to different strategies that are chosen by theparticipants in performing a given task. We referto these features as “inherent,” butwe acknowledgethat these can be “structurally inherent” (relatedto brain organization and connectivity) or “func-tionally inherent” (related to the cognitive state ofsubjects during the resting-state scan).The idea that brain connectivity can predict

activation has previously been reported for dif-ferentmodalities (22, 23), where diffusionMRI trac-tography was used (24) to measure connectivity.This study was limited to a number of visualcontrasts. More recently, resting-state connectivityhas been shown to be predictive of subjects’identity, in a way similar to a fingerprint (25).Rather than simply identifying subjects, our goalwas to predict the entire layout of brain activityfor each subject. Moreover, we also aim to pre-dict such layout of activity in a number of dif-ferent cognitive domains, from a single task-freescan, including in subjects that show patterns ofactivation that are different fromthegroupaverage(perhapsmost strikingly in right-lateralized subjectswhen the majority of training subjects are left-lateralized).There are important practical implications of

the proposed framework in basic research andtranslational neuroscience. It provides a methodfor inferring multiple individualized functionallocalizers based on a single resting-state scan.Such a tool could be used to investigate in detailthe response profiles of localized brain regionswithout the need to acquire often time-consumingtask localizers. Such a tool, if generalizable be-yond the young, healthy population that makesup the HCP database, could be used to inves-tigate functional regions in subjects who can-not perform tasks, such as infants or paralyzedpatients.

REFERENCES AND NOTES

1. M. Jenkinson, S. Smith, Med. Image Anal. 5, 143–156(2001).

2. J. Besle, R. M. Sánchez-Panchuelo, R. Bowtell, S. Francis,D. Schluppeck, J. Neurophysiol. 109, 2293–2305(2013).

3. F. McNab, T. Klingberg, Nat. Neurosci. 11, 103–107(2008).

4. I. Mukai et al., J. Neurosci. 27, 11401–11411(2007).

5. S. M. Tom, C. R. Fox, C. Trepel, R. A. Poldrack, Science 315,515–518 (2007).

6. G. S. Wig et al., Proc. Natl. Acad. Sci. U.S.A. 105, 18555–18560(2008).

7. R. Kanai, G. Rees, Nat. Rev. Neurosci. 12, 231–242(2011).

8. T. E. Behrens, H. Johansen-Berg, Philos. Trans. R. Soc. Lond. BBiol. Sci. 360, 903–911 (2005).

9. R. E. Passingham, K. E. Stephan, R. Kötter, Nat. Rev. Neurosci.3, 606–616 (2002).

10. B. T. T. Yeo et al., J. Neurophysiol. 106, 1125–1165(2011).

11. S. M. Smith et al., Proc. Natl. Acad. Sci. U.S.A. 106,13040–13045 (2009).

12. D. C. Van Essen et al., Neuroimage 80, 62–79(2013).

13. D. M. Barch et al., Neuroimage 80, 169–189(2013).

14. B. Fischl et al., Cereb. Cortex 18, 1973–1980(2008).

15. T. J. Andrews, S. D. Halpern, D. Purves, J. Neurosci. 17,2859–2868 (1997).

16. S. Knecht et al., Brain 123, 74–81 (2000).17. M. Thiebaut de Schotten et al., Science 309, 2226–2228

(2005).18. M. W. Cole, D. S. Bassett, J. D. Power, T. S. Braver,

S. E. Petersen, Neuron 83, 238–251 (2014).19. E. C. Robinson et al., Neuroimage 100, 414–426 (2014).20. N. K. Logothetis, B. A. Wandell, Annu. Rev. Physiol. 66,

735–769 (2004).21. K. J. Friston, Brain Connect. 1, 13–36 (2011).22. Z. M. Saygin et al., Nat. Neurosci. 15, 321–327 (2012).23. D. E. Osher et al., Cereb. Cortex 26, 1668–1683 (2016).24. P. J. Basser, S. Pajevic, C. Pierpaoli, J. Duda, A. Aldroubi, Magn.

Reson. Med. 44, 625–632 (2000).25. E. S. Finn et al., Nat. Neurosci. 18, 1664–1671 (2015).

ACKNOWLEDGMENTS

Data were provided by the Human Connectome Project, WU-MinnConsortium (Principal Investigators: D. Van Essen and K. Uğurbil;1U54MH091657) funded by the 16 NIH Institutes and Centersthat support the NIH Blueprint for Neuroscience Research; andby the McDonnell Center for Systems Neuroscience at WashingtonUniversity in St. Louis. The data are available for download at www.humanconnectome.org. Data from the Q3 release (September

2013) were used in this paper. All code available upon request fromthe corresponding author. Funding was provided by the U.K.Medical Research Council (MR/L009013/1 to S.J.), U.K.Engineering and Physical Sciences Research Council(EP/L023067/1 to S.J. and T.E.B.), James S. McDonnellFoundation (JSMF220020372 to T.E.B.), the Wellcome Trust(WT104765MA to T.E.B. and 098369/Z/12/Z to S.M.S.), and theNetherlands Organization for Scientific Research (NWO) (452-13-015 to R.B.M.). Author contributions were as follows: S.J. and I.T.built the model and performed the analyses. S.J., I.T., R.B.M., O.P.J.,S.M.S., and T.E.B. wrote the paper.

SUPPLEMENTARY MATERIALS

www.sciencemag.org/content/352/6282/216/suppl/DC1Materials and MethodsFigs. S1 to S23Tables S1 to S3Captions for Movies S1 to S7References (26–39)Movies S1 to S7

5 November 2015; accepted 29 February 201610.1126/science.aad8127

POLITICAL SCIENCE

Durably reducing transphobia:A field experiment ondoor-to-door canvassingDavid Broockman1* and Joshua Kalla2

Existing research depicts intergroup prejudices as deeply ingrained, requiring intenseintervention to lastingly reduce. Here, we show that a single approximately 10-minuteconversation encouraging actively taking the perspective of others can markedlyreduce prejudice for at least 3 months.We illustrate this potential with a door-to-doorcanvassing intervention in South Florida targeting antitransgender prejudice. Despitedeclines in homophobia, transphobia remains pervasive. For the intervention, 56canvassers went door to door encouraging active perspective-taking with 501 votersat voters’ doorsteps. A randomized trial found that these conversations substantiallyreduced transphobia, with decreases greater than Americans’ average decreasein homophobia from 1998 to 2012. These effects persisted for 3 months, andboth transgender and nontransgender canvassers were effective. The interventionalso increased support for a nondiscrimination law, even after exposingvoters to counterarguments.

Intergroup prejudice, defined broadly as neg-ative attitudes about an outgroup, is a rootcause of numerous adverse social, political,and health outcomes (1–3). Influential theo-ries depict intergroup prejudices as deeply

ingrained during childhood and highly resistantto change thereafter (4–6). Consistent with thesetheories, empirical research has found that du-rably reducing prejudice is challenging. Massmedia interventions and other brief stimuli usu-ally fail to reduce prejudiced attitudes (7) or haveonly temporary effects (8); lasting change ap-

pears to require intense intervention overmonths(9, 10). Rare are studies demonstrating prejudice-reduction interventions relatively brief in dura-tion yet proven to have lasting effects (4).Theories of active processing, however, suggest

a method for even brief interventions to durablychange attitudes. A recurring finding of labora-tory studies is that brief messages can durablychange individuals’ attitudes when individualsengage in active, effortful, processing (known as“System 2” processing) of those messages (11).These studies, conducted on other topics, raisethe possibility that brief interventions encour-aging active consideration of counter-prejudicialthoughts could produce lasting changes in attitudestowardanoutgroup.Perspective-taking, “imaginingthe world from another’s vantage point,” is one

220 8 APRIL 2016 • VOL 352 ISSUE 6282 sciencemag.org SCIENCE

1Graduate School of Business, Stanford University, Stanford, CA,USA. 2Department of Political Science, University of California,Berkeley, CA, USA.*Corresponding author. E-mail: dbroockman@stanford.edu

RESEARCH | REPORTSCorrected 11 April 2016; see full text.

on

Janu

ary

23, 2

017

http

://sc

ienc

e.sc

ienc

emag

.org

/D

ownl

oade

d fr

om

(6282), 216-220. [doi: 10.1126/science.aad8127]352Science and S. Jbabdi (April 7, 2016) I. Tavor, O. Parker Jones, R. B. Mars, S. M. Smith, T. E. Behrensduring task performanceTask-free MRI predicts individual differences in brain activity

 Editor's Summary

   

, this issue p. 216Scienceduring task-based fMRI.marker. Resting-state functional connectivity thus already contains the repertoire that is then expressedcognitive paradigms. This suggests that individual differences in many cognitive tasks are a stable trait were not performing any explicit task predicted differences in fMRI activation across a range of(fMRI) data from the Human Connectome Project. Brain activity in the ''resting'' state when subjects

applied computational models to functional magnetic resonance imaginget al.brain activity? Tavor We all differ in how we perceive, think, and act. What drives individual differences in evoked

Every brain is different

This copy is for your personal, non-commercial use only.

Article Tools

http://science.sciencemag.org/content/352/6282/216article tools: Visit the online version of this article to access the personalization and

Permissionshttp://www.sciencemag.org/about/permissions.dtlObtain information about reproducing this article:

is a registered trademark of AAAS. ScienceAdvancement of Science; all rights reserved. The title Avenue NW, Washington, DC 20005. Copyright 2016 by the American Association for thein December, by the American Association for the Advancement of Science, 1200 New York

(print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last weekScience

on

Janu

ary

23, 2

017

http

://sc

ienc

e.sc

ienc

emag

.org

/D

ownl

oade

d fr

om